Cellular/Molecular

Short-Term Memory in Networks of Dissociated CorticalNeurons

Mark R. Dranias, Han Ju, Ezhilarasan Rajaram, and Antonius M. J. VanDongenProgram in Neuroscience and Behavioral Disorders, Duke-NUS Graduate Medical School, Singapore 169857

Short-term memory refers to the ability to store small amounts of stimulus-specific information for a short period of time. It is supportedby both fading and hidden memory processes. Fading memory relies on recurrent activity patterns in a neuronal network, whereashidden memory is encoded using synaptic mechanisms, such as facilitation, which persist even when neurons fall silent. We have used anovel computational and optogenetic approach to investigate whether these same memory processes hypothesized to support patternrecognition and short-term memory in vivo, exist in vitro. Electrophysiological activity was recorded from primary cultures of dissoci-ated rat cortical neurons plated on multielectrode arrays. Cultures were transfected with ChannelRhodopsin-2 and optically stimulatedusing random dot stimuli. The pattern of neuronal activity resulting from this stimulation was analyzed using classification algorithmsthat enabled the identification of stimulus-specific memories. Fading memories for different stimuli, encoded in ongoing neural activity,persisted and could be distinguished from each other for as long as 1 s after stimulation was terminated. Hidden memories were detectedby altered responses of neurons to additional stimulation, and this effect persisted longer than 1 s. Interestingly, network bursts seem toeliminate hidden memories. These results are similar to those that have been reported from similar experiments in vivo and demonstratethat mechanisms of information processing and short-term memory can be studied using cultured neuronal networks, thereby settingthe stage for therapeutic applications using this platform.

IntroductionShort-term memory enables animals to preserve stimulus infor-mation across brief delays (Baddeley, 1996). The mechanismsunderlying these stimulus-memory processes involve transientchanges in the active and hidden states of neurons in a neuronalnetwork (Buonomano and Maass, 2009). This raises the questionwhether random networks formed by cultures of dissociated cor-tical neurons also have the ability to store stimulus informationfor short periods of time. To test whether the same mechanismsthat support short-term memory in vivo can encode stimulus-specific memories in vitro, we employed a novel combination ofoptogenetic and computational approaches.

Active or fading short-term memories are encoded in the ac-tion potential activity reverberating in a neuronal network. Thesesignals preserve stimulus identity information until the activitysubsides (Wang, 2001). Hidden memory traces store informationas a spatial pattern of synaptic adaptation, which persists even

when neurons fall silent (Buonomano and Maass, 2009). Thesehidden memories temporarily alter the timing or synaptic trans-mission of action potentials, and the presence of these memoriescan be revealed by the altered electrophysiological response of anetwork to additional stimulation (Mongillo et al., 2008).

Previous investigations into the information processing capa-bilities of neuronal networks formed from dissociated rat corticalneurons plated on multielectrode arrays (MEAs) have utilizedelectrical stimulation and shown that these networks can activelyrepresent stimulus information (Bakkum et al., 2008b). In thesestudies, the response of the network to electrical stimulation ischaracterized by two phases: an early phase of short latency actionpotentials (3–20 ms) that are directly evoked by stimulation, fol-lowed by a burst of firing activity across the network that peaksbetween 20 and 200 ms (Eytan and Marom, 2006; Gal et al., 2010;Kermany et al., 2010). These network bursts are characterized bya simultaneous, highly correlated firing of neurons across theculture, after which activity dissipates to baseline (Beggs andPlenz, 2003; Wagenaar et al., 2006a). A strong relationship existsbetween the distance of a neuron to the stimulating electrode andthe latency of its earliest response, enabling researchers to recon-struct the location of the stimulating electrode, often using la-tency information alone (Eytan et al., 2003; Dockendorf et al.,2009; Kermany et al., 2010).

The capacity of cultured neuronal networks for storing stim-ulus information or performing state-dependent computationsusing active and hidden processes has not been studied in vitro.We have therefore investigated the responses of cortical neuronsto complex stimuli in order to answer the following questionsregarding short-term memory mechanisms in randomly orga-

Received June 6, 2012; revised Nov. 11, 2012; accepted Dec. 7, 2012.Author contributions: M.R.D. and A.M.J.V. designed research; M.R.D. and E.R. performed research; M.R.D., H.J.,

E.R., and A.M.J.V. contributed unpublished reagents/analytic tools; M.R.D., H.J., and A.M.J.V. analyzed data; M.R.D.and A.M.J.V. wrote the paper.

This work was supported by United States National Science Foundation Grant ECCS-0925407 and an award fromthe Singapore Ministry of Health and A*STAR, the Agency for Science, Technology and Research. We thank HendrikaVanDongen for expert technical advice, Niamh Higgins for technical support, and Dr. Karl Deisseroth for the ChR2-EYFP plasmids.

The authors declare no competing financial interests.Correspondence should be addressed to Dr. Antonius M. J. VanDongen, Program in Neuroscience and

Behavioral Disorders, Duke-NUS Graduate Medical School, 8 College Road, Singapore 169857. E-mail:antonius.vandongen@duke-nus.edu.sg.

DOI:10.1523/JNEUROSCI.2718-12.2013Copyright © 2013 the authors 0270-6474/13/331940-14$15.00/0

1940 • The Journal of Neuroscience, January 30, 2013 • 33(5):1940 –1953

nized cultured neuronal networks: 1) What are the response dy-namics to the optical presentation of complex spatial stimuli? 2)How long do short-term memories of stimuli persist? 3) Cancultured neuronal networks store stimulus-specific informationusing both active and hidden memory mechanisms? Clarifyingthe mechanisms of short-term memory has implications fortreating memory disorders ranging from transient cognitive im-pairment to Alzheimer’s disease.

Materials and MethodsPrimary neuron cell culture. Cortical neurons obtained from rat E18 em-bryonic brain of either sex were dissociated and plated as described pre-viously (Van de Ven et al., 2005; Dranias et al., 2011) on MEA-containingculture dishes (Multi-Channel Systems). Two different electrode config-urations were used: a standard 60 electrode dish arranged in an 8 � 8 gridand a 252 electrode dish with electrodes arranged in a 16 � 16 grid. Priorto plating, MEAs were cleaned with detergent, rinsed, sterilized overnightunder UV lighting, and coated with poly-D-lysine and fibronectin (Invit-rogen). Cortical tissue from E18 rats is dissociated using papain and 500�l of DNase. The resulting cell suspension was rinsed, centrifuged, andthen the pellet was resuspended in NB-Active 4 cell medium (BrainBits),in some experiments supplemented with 10% FBS. Neurons slated to betransfected underwent electroporation at this point and then were platedonto the poly-D-lysine and fibronectin-coated MEA dishes. Cell culturewas maintained in NB-Active 4 cell medium to suppress astrocytegrowth.

Optogenetic transfection. Plasmid DNA encoding Channel Rhodopsin-2(ChR2) was a kind gift from Karl Deisseroth. ChR2 was fused to EYFP forvisualization, and mutations H134R and T159C were introduced toboost current (Nagel et al., 2005; Berndt et al., 2011). The plasmid in-cluded a cytomegalovirus promoter. Two transfection methods wereused: electroporation (Amaxa nucleofector II kit, Lonza) performed be-fore plating, or lipofection (Lipofectamine 2000, Invitrogen) at least 24 hbefore recordings on 7–14 days in vitro (DIV). ChR2-EYFP expressionwas observed within 24 h after transfection.

MEA recordings. Extracellular electrophysiological recordings of neu-rons were made from 60 and 252 electrode MEA dishes using theMEA1060 and USB-MEA256 hardware systems, respectively (MultiChannel Systems). Recordings were performed on an antivibration tableand in a Faraday cage. Signals were amplified and sent to a personalcomputer using a PCI-based data acquisition card for the MEA1060system or via high speed USB for the USB-MEA256. During experimen-tal recordings, the cell culture medium of each MEA dish was replaced

with Dulbecco’s phosphate-buffered saline so-lution containing glucose and pyruvate(Sigma).

Data analysis. MC_Rack software (Multi-channel Systems) was used to acquire electro-physiological signals in real-time. Extracellularsignals were high pass filtered at 300 Hz andlow pass filtered at 3 kHz with second orderButterworth filters. Action potentials were de-tected using a voltage threshold rule whosevalue was determined by the user for each dishbased on the amount of channel noise. Whenan action potential was detected, the timing ofthe spike was recorded together with a 2 mssample of the waveform, sampled at 120 kHz.

Optical stimulus presentation and imaging.The MEA system was mounted on an invertedmicroscope (Eclipse Ti, Nikon) during record-ings. A 25 mW 488 nm laser (Spectra-Physics)beam was projected through an acousto-optictunable filter (AA Opto-Electronic), which al-lows the laser light to be controlled by a TTLpulse generated by the personal computer. Thebeam was optically expanded, polarized, andprojected at a 45° angle onto a reflective LCoSSpatial Light Modulator microdisplay (SLM,

Holoeye Photonics AG) with a resolution of 1920 � 1200 pixels and a 60Hz refresh rate. The status of the pixels is controlled using the DVI port ofa personal computer allowing images to be generated in MATLAB. Bluelight patterns reflecting off the SLM were passed through a second polar-izing filter and projected onto the neuronal network growing on top ofthe MEA. Fluorescent and bright-field images were captured from theMEA dishes via a cooled CCD camera (Orca, Hamamatsu) and analyzedusing NIS Elements software (Nikon). This enabled verification of opti-cal stimuli and visual inspection of electrophysiologically active chan-nels. Figure 1 illustrates the optical path of the custom laser projectionsystem. Custom drivers and MATLAB software were used to generate thestimulus patterns, control the onset and offset of optical stimuli, andgenerate TTL pulses to synchronize electrophysiological recordings withthe presentation of laser stimuli.

Stimulus material. Two sets of stimuli were used in these experiments.The first set of stimuli consisted of a sequence of lighted squares from a2 � 2 grid or a 5 � 5 grid. These stimuli were presented to assess theefficacy of transfection. The grid occupied a region of �1.5 mm 2, and theindividual squares in the grid were of variable size (0.3– 0.75 mm 2). Inthe experiments testing short-term memory for complex stimuli, ran-dom dot stimuli were presented. The random dot stimuli were con-structed from 18 to 22 randomly positioned squares on a 10 � 10 grid.When projected onto the MEA dish, the 10 � 10 grid occupied an area of�1.25 mm 2. Effective light intensity was measured as 0.1 mW/mm 2. Thelight was projected through a low-power objective (Fig. 1), which al-lowed us to stimulate neurons located anywhere in a 1.5 mm 2 field ofview. A consequence of this approach is that relatively long (100 –200 ms)optical stimuli are required to elicit light responses. At this intensity, themean latency to the first recorded neural action potential was 165 � 4.7ms (N � 115 neurons).

Preprocessing of network activity. Electrophysiological data were im-ported from MC_Rack into MATLAB using the Neuroshare API library(www.neuroshare.org). Data were grouped into single trial spike timingvectors. Before analysis by machine classifiers, spike timing data werecounted and discretely resampled into individual spike times (liquid-state machine [LSM]) or nonoverlapping 5, 50, or 500 ms bins, depend-ing on whether the data were to be analyzed as a time series or a singletrial. Only spike timing data from electrodes reporting at least 1 spike perminute were included for analysis.

Machine classification of unit responses. Biologically plausible classifi-cation algorithms were implemented in MATLAB to analyze single trialdata. The simplest and most plausible anatomy for such a model is areadout neuron with tunable weights that allow the neuron to make

Figure 1. Optical path for the custom laser projection system. The beam of a 25 mW 488 nm laser passes through an acousto-optic tunable filter onto a reflective LCoS SLM. Blue light patterns reflecting off the SLM were projected onto the neuronal networkusing the lenses as shown. Components enclosed by the dotted outline are located inside the inverted microscope.

Dranias et al. • Short-Term Memory J. Neurosci., January 30, 2013 • 33(5):1940 –1953 • 1941

binary category membership decisions. Two algorithms were chosen totune these weights: a support vector machine (SVM) or LSM algorithm.All algorithms have distinct training and testing phases (Haykin, 1999).During the training phase, a sample of data is taken to tune the weightsusing the appropriate algorithm. In the testing phase, no new learningtakes place and the trained neural network is used to classify the remain-ing testing data. This assays the generalizability of the categories learnedduring the training phase. Data were generally divided into 70 –30%training and testing trials. Classifications were retrained and retested5–50 times choosing random subsets. The accuracy of the classifier istaken as the mean percentage correctly classified test trials. In multiclassclassification, each readout neuron is trained to make a “one versus all”category membership decision. Multiclass classifications require multi-ple readout neurons arranged in a winner-take-all competition to deter-mine category membership.

The SVM was implemented using the online resource libsvm (Changand Lin, 2011; http://www.csie.ntu.edu.tw/�cjlin/libsvm). Data were in-put as the mean activity in a 50, 100, or 250 ms time window. Thedimensionality of the inputs and weights is equal to the number of activeelectrodes. As category membership decisions are binary outputs, thereadout neuron was simplified and represented by just its synaptic weightmatrix.

The LSM was used because living neural networks are known to usetemporal signals to encode stimulus information (Kermany et al., 2010)and the LSM is sensitive to temporal patterns (Maass et al., 2002). TheLSM was simulated using CSIM software (Natschlager and Markram,2003). The LSM is a form of filtered classifier that nonlinearly separatespatterns, increasing the probability that any two different inputs to theLSM will be linearly separable (Maass et al., 2002). The LSM consists ofan input layer, a liquid filter or reservoir of randomly connected neurons,and a linear classifier that separates the signals that propagate through theneural reservoir. MEA data were used to drive the input layer to the liquidfilter. To preserve spatial information of each electrode, the input layerconsists of 8 � 8 neurons arranged equivalent to the MEA layout. Eachinput neuron receives recorded spikes from one electrode and propagatesthe input into the reservoir of simulated neurons that contains 8 � 8 � 8neurons in a 3D space, with 80% of the neurons being excitatory and theremaining 20% being inhibitory. The Multiple-timescale AdaptiveThreshold neuron model was used, with a firing threshold of 14 mV,resting membrane potential 0 mV, 30 ms time constant, and a refractory

period of 2 ms (Kobayashi et al., 2009). Parameters for the thresholdadaptation are set to: �1 � 20 mV, �2 � 5 mV, �1 � 2 ms, and �2 � 40 ms.Connection probability between two neurons was based on the Euclid-ean distance:

P�D� � C*exp� � D2�a,b�

�2 �,

where D(a,b) represents the Euclidean distance between neurons a and b,� � 5 and C � 0.5 for the input connections, and � � 7 and C � 0.6 forconnections in the reservoir. Values were chosen after parametric simu-lations to optimize classification on a test set of data. All the synapticweights were randomly initialized following a � distribution. Input con-nections are static, whereas the connections in the reservoir incorporateshort-term plasticity with CSIM default parameters. A readout neuron isconnected to all the neurons in the reservoir to make classification deci-sions. Its synaptic weights are optimized by Fisher’s linear discriminant.For multiclass classification, multiple readout neurons were created and

Figure 2. Stimulus presentation protocol. A, Stimulus-specific memory trials. Thirty staticpatterns of randomly generated dots were presented to the neuronal network for 100 –200 ms.Neuronal activity is recorded for 500 ms before stimulus onset and for 1–2 s after stimulus onset.B, Cue-Probe Short-term memory trials. Two stimuli were presented sequentially to the neu-ronal network: a cue stimulus followed by a probe stimulus after a short delay. The cue stimuluscould be one of two (or four) random dot stimuli, whereas the probe stimulus was fixed acrosstrials. Each stimulus was presented for 100 –200 ms. Cue and probe stimuli are separated by avariable delay, ranging from 200 ms to 2 s.

Figure 3. ChR2 transfection. A, Light microscopic image of primary neuronal culture at DIV9on a 60 electrode MEA transfected with plasmid DNA encoding Channel Rhodopsin-2 fused toYellow Fluorescent Protein (ChR2-YFP). B, Image taken using a 4� objective and 510 nmexcitation light to visualize ChR2-YFP expression. Cell count indicates approximately 2% trans-fection efficiency.

1942 • J. Neurosci., January 30, 2013 • 33(5):1940 –1953 Dranias et al. • Short-Term Memory

a winner-take-all rule was applied using the average activity of the read-out neuron over the trial. A new Liquid State Filter topology was gener-ated each time the network was retrained and tested.

Classifier information time series. A time series was constructed to mea-sure how stimulus information varies over time. The time series wereconstructed using multiple, independent LSMs or SVMs. Each LSM orSVM analyzed the spike times or mean firing rates from a small timewindow. For the LSM and SVM network, spike times were segmentedinto 100 or 50 ms bins. An independent LSM or SVM is dedicated to eachbin and trained to classify the neuronal response on this interval. Usingan array of independent classifiers allowed the neuronal response data tobe replaced with a time series of data points, whose value represents theamount of information available at that moment for predicting whichstimulus was presented at the start of the trial. The amount of availableinformation was determined using the percentage of correctly classifiedtesting trials. The classifiers were implemented independently, and thisprovides a lower limit on the amount of information (Nikolic et al.,2009). Unless otherwise stated, the performance measure is the classifi-cation accuracy on the testing set.

Sources of classification information. The SVM uses two basic sources ofinformation to perform classification: the spatial pattern of activity andthe magnitude of overall activity. The spatial pattern of activity consistsof an activity vector that is the combination of units that are activated bystimulus presentation. The LSM performs classifications primarily rely-ing on three sources of information: spatial pattern of activation (spatialinformation), the magnitude of overall activity (network spike rate), andthe temporal pattern of spiking inputs (temporal information). The tem-poral encoding of a stimulus relates to sequences of spikes, spike latenciesor variations in firing rate that are stimulus-specific; these properties areextracted by the reservoir functionality of the liquid filter. The contribu-

tion of each of these sources of information toclassification accuracy was measured by filter-ing out the spatial and/or temporal informa-tion before training and testing. For the LSM,the role temporal information plays can be un-covered by comparing the rate of classificationbefore and after randomizing the spike times.Spike times are randomly drawn from a uni-form distribution over the same interval (50 or100 ms bins). For both SVM and LSM, the roleof spatial information can be uncovered bycomparing the rate of classification before andafter the pooling of active units. Pooling of ac-tive units eliminates all information aroundthe spatial pattern of activation. After filteringout spatial information for the SVM and bothspatial and temporal information for the LSM,the amount of information encoded by overallmean firing rate can be assessed.

Transfection assessment. These experimentswere used to determine whether dishes had ef-fective transfection. Four square stimuli from a2 � 2 grid were presented sequentially, in aclockwise order. Each square stimulus was typ-ically presented for between 150 and 250 msand poststimulus response was monitored for500 ms to 1.5 s. If a culture showed a reliableresponse to one or more squares, it was used inadditional experiments. Dishes used in furtherexperiments ranged from DIV8 to DIV20.

Experimental protocols. Two basic experi-mental protocols were followed as illustrated inFigure 2. Stimulus-specific memory trialssought to determine the capacity and durationof short-term memories in the MEA dish. Upto 30 different single random dot stimuli werepresented for 100 –200 ms and the poststimu-lus response was monitored for 1.5–2.5 s. Thestimuli were presented in pseudorandom or-der, usually 800 –1200 trials per dish. During

experiments where all 30 stimuli are presented, each stimulus is repeated�30 times; whereas during other experiments where only four stimuliare presented, there were �200 repetitions each. Spike times from elec-trophysiological recordings of single units were collected into 250 msbins. Multiclass classification was used to analyze the network responseto discover how many different stimuli can be discriminated using net-work responses. To determine the duration of stimulus-specific short-term memory, classification between 4 random dot stimuli was used. Thespecificity and duration of the memory were determined by a time seriesconstructed using an array of LSMs or SVMs.

Short-term memory/delayed response trials sought to determinewhether the dish can support short-term memories that could be used tosolve a task. In these trials, two stimuli (a cue and a probe stimulus) arepresented sequentially, separated by a pause. If memories of the cuestimulus influence the processing of the probe stimuli, then the networkcould use this memory to select actions or responses. The cue stimuluswas either one of two or one of four random dot stimuli presented for100 –200 ms (Fig. 2B). The probe random dot stimulus was presented200 –2000 ms later for 100 –200 ms. The amount of cue-specific informa-tion was measured using a time-series constructed by an array of LSMs orSVMs.

Statistical methods. Random label shuffling was used to measure thesignificance of the classification accuracies. Label shuffling was repeated100 –200 times, and the best and worst classification rates (95th percen-tile) were recorded for each classifier to establish the significance of thecorrect classification rates. Bias relating to the unfair sampling of trainingand testing vectors was controlled by retraining and testing the classifiers50 times. For each repeated training and testing, a different set of trainingand testing data were randomly selected from the experimental data. The

Figure 4. Unit responses encode spatial stimuli. A, Left, Square stimuli were projected onto a DIV14 network for 200 ms usinga 488 nm laser to excite ChR2 at different locations. Right, Images of the culture taken while illuminated by the two laser stimuli.Graphs in Figure 4B are based on an analysis of data from a single electrode at position (6,2) shown circled. B, C, Raster plots ofaction potentials recorded extracellularly from electrode position (6,2). Responses from 45 trials are shown. Stimulus onset andoffset are indicated by green and red vertical lines, respectively. C, D, For Stimulus 1 (C) and Stimulus 2 (D), the average latencieswere 92 and 156 ms.

Dranias et al. • Short-Term Memory J. Neurosci., January 30, 2013 • 33(5):1940 –1953 • 1943

mean and its SE are reported from these 50repeated training and testing episodes. Thefalse-positive rate was used in measuring thestimulus representation capacity of the cul-tured neuronal networks. The correlation be-tween the distance of an electrode from thestimulus and the latency of its action potentialswas computed. The distance of a neuron to thestimulus was measured as the distance of therecording electrode to the stimulus center. Dis-tance was measured in electrode units. The la-tency was taken as the mean latency fromstimulus onset until the first spike recordedfrom that electrode. Neurons were includedin the analysis if they responded reliably totwo of the square stimuli (20% of presenta-tions) inside a 300 ms window beginning atstimulus onset. When comparing two condi-tions, the change in classification accuracywas calculated by recording the percentagechange after the chance rate of correct clas-sification was subtracted out from the signalsunder consideration.

ResultsPrimary cultures of dissociated E18 ratcortical neurons were grown on an MEAand transfected with a plasmid encodingChR2-EYFP. The MEAs were plated with10 5 cells, and the percentage of cells ex-pressing ChR2-EYFP averaged �2% (Fig.3). Experiments were performed on cul-tures 7–14 DIV, when on average 60% ofthe electrodes revealed spontaneous neu-ronal activity and 10 –15% of the elec-trodes were recording signals that couldbe reliably produced by presenting opticalstimuli. The method of transfection influ-enced the neuronal response; and while 28 MEAs had reliableresponses to the presentation of monochrome square stimuli, thelongest duration responses were recorded from MEAs trans-fected using electroporation (n � 8).

Representation of optically presented stimuli in livingneuronal networksFirst, we investigated how single units responded to simple spatialstimuli. Five MEA dishes were tested using square stimuli se-quentially presented from a 5 � 5 grid. The square stimuli were�0.6 � 0.6 mm. Ninety-five units responded to the presentationof these squares; 13% of the units responded to only one stimu-lus, while the remainder responded to more than one stimulus.Only 49% of units responded when a stimulus was presenteddirectly over the recording electrode. The remaining units (51%)could only be activated by illuminating regions more distal fromthe recording site. Distances from which stimuli were able toelicit responses ranged from directly overhead to as far as 2.5 mmaway.

The latency at which single neurons responded to spatial stim-uli did not increase with stimulus distance. Figure 4 displays datafrom experiments where single squares from a 2 � 2 grid wereprojected onto 10 MEA dishes (squares were 0.9 � 0.9 mm). Nocorrelation was found between the latency to firing of a singleunit and the distance of the stimulus from the site of the record-ing electrode (r � 0.10, p � 0.16, 119 neurons, 10 MEAs). Whensingle squares from a higher resolution 5 � 5 grid were used, the

latency of units was found to be weakly but inversely correlated tothe distance of the stimulus from the recording electrode (r ��0.11, p � 0.05, 95 neurons, 5 MEAs). Even when neurons thatresponded to multiple stimuli were examined individually, theaverage correlation between latency and distance for each unitwas negligible (r � �0.01, 82 neurons, 5 MEAs). Although thereis no systematic relationship between response latency and elec-trode distance, the latency of the responses of individual units wasfairly deterministic and predictable (average coefficient of varia-tion, 0.265). The average overall latency for these square stimuliwas 137 � 63 ms.

Next, we investigated the dynamics of neuronal responses tomore complex spatial stimuli. Random dot stimuli were pro-jected onto the ChR2-expressing neurons to test the ability of thenetworks to discriminate complex stimuli. Optical stimulationwith random dot patterns or squares led to an initial networkresponse that is characterized by the selective activation of a fewunits with an average latency of 165 ms after stimulus onset(range, 53–239 ms; N � 115 neurons, 13 MEAs). The informa-tion carried by this initial response decays by 50% after 122 � 36ms (n � 7 MEAs). On some trials, this initial response is followedby a synchronized burst of network activity. A typical networkburst lasts between 200 and 600 ms, after which there is a returnto baseline unit activity. If no burst occurs, the initial stimulus-elicited response will continue to dissipate until baseline responselevels are achieved after 200 –300 ms. The occurrence of a net-work burst is dependent on past network activity and the identity

Figure 5. Spike Raster plots recorded from single units in response to optical stimulation. Optical stimuli were presentedfor 100 ms. Each row represents one trial. Spike times are represented as circles, which are color coded according toelectrode number. Trials are aligned to the first spike time from the cyan unit and then sorted by the spike time recordedfrom the unit, which fired second on most trials. A, Stimulus 1 presented to a DIV14 MEA. An obvious alternating spikingactivity is observed. The average latency of the cyan unit is 114 ms after stimulus onset (14 ms after stimulus offset). x-axisindicates time from earliest latency spike. B, Response of same network to a second random dot stimulus. Alternatingactivation in response to optical presentation of a random dot stimulus is seen. The average latency of the cyan unit is 126ms after stimulus onset (26 ms after stimulus offset).

1944 • J. Neurosci., January 30, 2013 • 33(5):1940 –1953 Dranias et al. • Short-Term Memory

of the stimulus. Some stimuli elicit a network burst on 100% oftrials, whereas others seemingly never elicit a network burst. Dur-ing the initial 200 ms after optical stimulation, the response ofsingle units does not follow a simple profile of activation followedby passive decay. Figure 5 plots the responses of single units froma DIV9 MEA. The units are color coded, and from Figure 5 it canbe seen that most units start responding after the optical stimulushas terminated and the responses recorded at electrodes reveal acomplex interplay between units that is typical of the activationobserved in recurrent networks (Schrader et al., 2008).

Representational capacity for stimulus informationNetwork responses could reliably encode complex spatial stimuliconsisting of �20 randomly spaced dots. Thirty different ran-dom dot stimuli were presented to 8 cultured neuronal networksto measure how many stimuli can be reliably discriminated usingelectrophysiological recordings of network activity. The stimuliwere presented in pseudorandom order using the protocolshown in Figure 2A. For individual dishes, the mean rate of cor-rect classification measured on the testing set ranged from 0.14 to0.5. Classifications were performed on a 250 ms bin of networkactivity that began at stimulus termination. Subsequent bins wereexcluded from the analysis as classification accuracy fell precipi-tously thereafter. Figure 6A shows that, on average, 28 of the 30stimuli were classified at rates above chance when an SVM clas-sifier was used, whereas 29 of 30 stimuli were classified at ratesabove chance when the LSM was used (n � 8 MEAs).

Machine classifiers can address the question of how stimulusidentity information is encoded by the network. Stimuli can beencoded in the temporal patterns of spiking activity, overall spikecounts, or the spatial patterns of active channels (Kermany et al.,2010). How stimulus identity information is encoded can be an-alyzed by filtering out temporal and/or spatial information priorto classification. We have employed a liquid state machine (LSM)classifier for this purpose (see Materials and Methods). Figure 6Bshows that both temporal and spatial information plays a signif-icant role in stimulus encoding. Eliminating temporal structureby randomizing the spike times reduces classification accuracy by30%. When the spatial pattern of channel activation is eliminatedby pooling the data, there is a 69% reduction in classificationaccuracy. When both spatial and temporal information is elimi-nated, only the overall spike count can be used for classification.This results in a 94% drop in classification accuracy (almost tochance levels). In this case, only 5 stimuli are discriminated atrates above chance. These results indicate that the spatial patternof active units is the most important source of information fordiscriminating these spatial stimuli, although temporal informa-tion still makes a significant contribution.

Fading memory processesTo measure the duration of active stimulus memories, four ran-dom dot patterns were presented in pseudorandom order onto18 networks, using the protocol shown in Figure 2A. The dura-tion of active or fading memories was defined as the time thatelapsed between the end of the optical stimulus and the timewhen stimulus-specific information was no longer detectable inthe spiking activity of the cultured neuronal networks. When anSVM was used to classify these stimuli, the average stimulusmemory persisted for 354 � 70 ms (range, 14 –1480 ms) after theoptical stimulus was terminated. Analyzing the data from fourdishes using the multiclass LSM revealed that stimulus-specificinformation persisted in these networks for an average of 675 msafter stimulus (range, 150 –1200 ms). Figure 7 shows how classi-

fication accuracy varies over time, for one of the networks ana-lyzed using a multiclass LSM.

Synchronized network bursting is a universal property of cul-tured neuronal networks (Maheswaranathan et al., 2012). Wetherefore investigated whether the duration of active stimulusmemories differs during trials with bursting and nonburstingnetwork responses. The responses of the neuronal network to thesame stimulus were not simple and homogeneous: some presen-tations of a stimulus would elicit a network burst, whereas otherswould not. Figure 8 displays data from the same neuronal net-work as shown in Figure 7, except that the trials are sorted intobursting and nonbursting trials. When these bursting and non-bursting trials are analyzed separately using a multiclass LSM,stimulus information is observed to persist for different dura-tions. On nonbursting trials, the mean rate of correct classifica-tion was 28% above chance, but the duration of the signal was

Figure 6. Representational capacity and encoding of stimulus information. Classifiers weretrained on MEA data to discriminate neuronal responses to 30 different random dot stimuli. A,Comparison of different methods for classifying stimulus-elicited responses: multiclass SVMclassifier (hyphenated lines) versus multiclass LSM classifier (solid lines). Data from eight net-works were analyzed using eight separate SVM and LSM classifiers. Stimuli are ranked in orderof classification accuracy. Classifiers were trained using a 250 ms window. Blue lines indicate themean correct classification rate across all 8 networks; and red lines indicate the rate of false-positive classification associated with each stimulus. When the red line is above chance, this canindicate that the classifier has a measurable lack of specificity. Shaded areas represent the SEM.The chance rate of correct classification was 3.3% (hyphenated black line). Data points forindividual networks were obtained by dividing the data into 70%/30% training/testing setsand then taking the mean of 50 epochs of retraining and retesting. B, Temporal and spatialinformation were filtered out to determine how much stimulus information is encoded by thespatial pattern of active units or by temporal variations in firing rate. A solid blue line indicatesthe control rate of correct classification of the LSM for 30 different random dot stimuli. Whentemporal information is eliminated by replacing actual spike times with an equal number ofrandomized spike times, a slight drop in classification accuracy is seen (hyphenated blue lines).When information about electrode position is eliminated by pooling the data across electrodesinto a single channel, there is a very large drop in classification accuracy (dotted blue lines).Other conventions as in Figure 6A.

Dranias et al. • Short-Term Memory J. Neurosci., January 30, 2013 • 33(5):1940 –1953 • 1945

only 120 ms (n � 7 MEAs). On bursting trials, the mean rate ofcorrect classification was 14% above chance, whereas stimulus-related information could be correctly classified for 400 ms (n �7 MEAs). The multiclass LSM indicates that the initial networkresponse to the optical stimuli contains the highest amount ofstimulus-specific information for both trial types (Fig. 8B,C).However, network bursts contain a small but significant amountof stimulus-specific information. This indicates that networkbursts are capable of encoding information about stimulusidentity.

The mechanisms for encoding stimulus information vary dur-ing bursting and nonbursting trials. Filtering out spatiotemporalfeatures of the data before classification reveals that temporalspike patterns play a more important role representing stimulusinformation on bursting trials than on nonbursting trials. In con-trast, the spatial pattern of active channels is more important fordiscriminating stimuli during nonbursting than bursting trials.Although significant, overall spike count was the least informa-tive source of information on stimulus identity. When spatialinformation was eliminated by pooling, the accuracy of classifi-cation was reduced by 45% during nonbursting trials and 24.8%during bursting trials. When temporal information was elimi-nated by randomizing spike times, the accuracy of classification

was reduced by 9.7% during nonbursting trials and 19.7% duringbursting trials (n � 7 MEAs). This sensitivity to temporal infor-mation on bursting trials seems to correspond to stimulus-related differences in the time at which a network burst isinitiated. Finally, when both spatial and temporal informationwas eliminated and only overall spike counts were available forclassification, classification accuracy was reduced by similar de-gree for both nonbursting (66%) and bursting trials (62.5%).Although the spatial pattern of active units is important for bothbursting and nonbursting trials, it is not static (e.g., Figure 5). Tomeasure whether the subset of active units that is activated by astimulus is preserved across the fading memory, an SVM wastrained on a 50 ms window of network activity that began justafter stimulus offset. This SVM was then used to classify networkresponses recorded from successive 50 ms time bins. Generaliza-tion was extremely weak and limited to adjacent bins (n � 7MEAs).

Hidden memory processesWe have investigated the presence of hidden memory pro-cesses using sequentially presented stimuli according to thecue-probe protocol shown in Figure 2B. In the cue-probe task,two stimuli are presented sequentially and a classifier is used

Figure 7. Multiclass classification of neuronal responses to four random dot stimuli. A, Time series indicates the duration of stimulus-specific memories by charting how stimulus classificationinformation varies in time. Data are from a DIV11 neuronal network. Time series is composed of points constructed from an array of independent multiclass LSM classifiers that analyze nonover-lapping 50 ms time windows of the spike timing data. The multiclass LSM classifiers were trained to discriminate 4 stimuli; 70% of the data was used for training and the remaining 30% for testing.The mean rate of correct classification during training and testing trials is plotted in red and blue, respectively. Shaded regions indicate the SEM across 50 epochs of retraining and retesting. Chancerate of classification for 4 stimuli is 25% (hyphenated black line). Dotted lines indicate the 95% confidence intervals for random label reassignment. y-axis indicates the proportion of correctlyclassified trials; and x-axis, time in seconds. B, Average spike rate per trial (Hz) across all channels in the network. A network burst tends to occur in the vicinity of 300 ms after stimulus presentation.C, The four random dot stimuli used in these experiments. D, Heatmap of average spike rates per electrode channel generated in response to presenting the four experimental stimuli. Spike rateswere calculated in 20 ms bins; color values range from 0 to 2 Hz. Data are from 45 electrodes.

1946 • J. Neurosci., January 30, 2013 • 33(5):1940 –1953 Dranias et al. • Short-Term Memory

to analyze the response of the network to the probe stimulusand determine whether probe-evoked activity is the same onevery trial or whether it changes based on the cue stimulus thatpreceded it. If the classifier can discriminate which cue pre-

ceded the probe at a rate above chance,this constitutes a “one-back” memory(Nikolic et al., 2009). Ten in vitro net-works were tested using this protocol;and in 6 of 10 networks, such a one-backmemory was seen: the response of thenetwork to the probe stimulus varies ac-cording to the cue stimulus that pre-cedes it. For the dishes with such one-back memory, analysis of a 200 ms timewindow after probe presentation showsthat there was a recovery of cue-relatedinformation relative to control. On tri-als where the probe was presented, theidentity of the cue could be correctlydiscriminated at a rate of 22 � 6% abovechance. In comparison, on trials whereno probe stimulus was presented, theidentity of the cue was correctly classi-fied at a rate of 4 � 2% above chance.Presentation of a probe stimulus unre-lated to the cue therefore increased clas-sification accuracy in each case. Figure 9shows several cases where cue related in-formation peaks during cue presenta-tion, dissipates, and then recovers when

a second, unrelated, probe stimulus is presented. The delaybetween cue and probe stimuli varied between 200 and 1500ms in these experiments.

Figure 9. Response to probe stimulus is modulated by the identity of previous cue stimuli. A–D, Four time series are shown,revealing the amount of cue-related information in the network responses. For each experiment, one of two different random dotcue stimuli (“X/Y”) was presented. On experimental trials (red), the cue stimulus was followed by a probe stimulus (“Z”). On controltrials (blue), the probe stimulus was not presented. Independent binary LSM classifiers analyzed nonoverlapping 50 ms timewindows of the spike timing data; 70% of the data was used for training and the remaining 30% for testing. The age of the culturesand pause between cue and probe were as follows: A, DIV11, 800 ms; B, DIV9, 400 ms; C, DIV8, 200 ms; D, DIV9, 1 s.

Figure 8. Encoding of stimulus information on bursting and nonbursting trials. A, Raster plot of action potentials. The times at which action potentials were recorded from different electrodesin response to the presentation of the four different experimental stimuli were pooled across the network and are shown plotted as black hash marks. Dish is the same as Figure 6. y-axis indicatesindividual trials. B, Encoding of stimulus information on nonbursting trials. Hyphenated black line indicates accuracy of classification of testing data by LSM. Thin solid blue line indicates classificationaccuracy when temporal information is filtered out by randomizing spike times (‘jittering’), and classification is performed relying only on overall firing rate and the pattern of active electrodes. Solidgreen line indicates classification accuracy when spatial information is eliminated by pooling spikes across all electrodes into a single channel, leaving temporal information and overall firing rateto guide classification. Dotted red line indicates classification rate based only on overall firing rate; both spatial and temporal data are filtered out of data before LSM classification. The time seriesconsists of 100 ms bins. Chance rate of classification is 0.25 (solid gray line). C, Encoding of stimulus information on bursting trials. Conventions as in B.

Dranias et al. • Short-Term Memory J. Neurosci., January 30, 2013 • 33(5):1940 –1953 • 1947

The duration and capacity of short-term memory during cue-probe experi-ments were much larger than what wouldbe predicted from the analysis of active(fading) network responses. The durationof the one-back memory appears to ter-minate at some point after 1.5 s. Figure 10shows data from an experiment where thedelay between cue and probe stimuli was1.2 s. Short-term memory could not bedetected at delays of 2 s or longer. At ashorter delay up to 1400 ms, the responseto the probe stimulus was significantlymodulated by the identity of the cue stim-ulus that preceded it.

It is possible that this one-back memorydoes not encode stimulus-specific informa-tion and simply reflects the influence of ha-bituation on an all-or-none network burst.To investigate this question, we performedfour experiments where we used 4 differentcue stimuli instead of the usual 2. A strongone-back memory was observed in 2 of 4dishes. Figure 11A charts the amount of cue-related information recorded from a net-work whose response to a probe stimuluswas modulated by all 4 preceding cues, re-sulting in a classification accuracy rate of90%. Filtering the classifier data indicatesthat, for this dish, cue-related informationwas primarily encoded by the spatial config-uration of active units.

To determine whether additional cue-related information could be revivedwhen a second probe stimulus was pre-sented, a cue-probe-probe stimulus se-quence was presented to four dishes. Twoof four dishes showed a one-back mem-ory, where the probe presentation elicitedan increase in the amount of cue-relatedinformation. However, in none of thedishes did the presentation of a secondprobe lead to a similar recovery of cue-related information. Figure 11B shows acase where one-back but no two-backmemory was seen.

Figure 10. Response to probe stimulus is modulated by identity of cue stimuli. A, Time series charting cue-related informationduring an experiment with a DIV11 culture. Cues are presented for 100 ms, the delay is 1.2 s, and the probe is presented for 100 ms.All stimuli are random dot patterns. The amount of cue-related information is measured using a time series constructed from anarray of binary LSM classifiers. Presentation of the probe stimulus causes a recovery of cue-related information, and that

4

information persists at levels above chance for a further 1.2 s.Classifiers were trained and tested on 492 trials (246 trials perstimulus). Chance rate of correct classification is 50% (hyphen-ated black line). Other conventions are as in Figure 6. B, Net-work activity in response to Stimulus 1 (blue) and Stimulus 2(red). Lines indicate the mean spike rate per trial (Hz) across allchannels in the network. A network burst tends to occur in thevicinity of 300 ms after stimulus presentation. C, Raster plot ofaction potentials. Responses to Stimulus 1 and Stimulus 2 areshown on the left and right, respectively. Vertical lines indicatestimulus onset and termination. D, Heatmapofmeanspikeratesfor each electrode. Spike rates were calculated in 20 ms bins; colorvalues range from 0 to 2 Hz. Other conventions as in Figure 6.

1948 • J. Neurosci., January 30, 2013 • 33(5):1940 –1953 Dranias et al. • Short-Term Memory

Short-term memory relies on both active andhidden mechanismsTo investigate whether short-term memory relies on spiking ac-tivity or hidden memory processes to bridge the delay betweencue and probe, responses from the previous experiment weresorted into bursting and nonbursting trials (Fig. 12). Two neuro-

nal networks showed that a short-termmemory with lumped trials also showed ashort-term memory on nonbursting tri-als. From the raster plots in Figure 12 itcan be seen that no activity bridges the gapbetween cue and probe presentations.Nevertheless, the network responds dif-ferently to the probe depending on whichstimulus preceded it. In Figure 12A, thenetwork is presented with a random dotstimulus “X.” When “X” is followed by theprobe stimulus “Z,” there is a short re-sponse to the stimulus but no networkburst. In Figure 11B, the same probe stim-ulus has been preceded by the other ran-dom dot stimulus “Y.” This time, however,the response of the network to the probe isvery different: there is a network burst. Forthis dish, the cue stimulus appears to alterthe probability that the probe will elicit anetwork burst. Figure 12C, D shows the re-sponses of a second dish with very differentdynamics. Although this dish had sparse ac-tivation on nonbursting trials, it alsoshowed a short-term memory. In this case,the probe stimulus only activates 3–4 unitsin the multielectrode array and “short-term” memory appears to be a result of cue-dependent adaptation in two of the threeunits. Analysis of how the hidden mem-ory was encoded in the network activityrevealed that the information was pri-marily encoded in overall firing rate,whether the units engage in a networkburst or not.

Hidden memory was disrupted duringbursting trials. When the same dishes thatdisplayed hidden memory during non-bursting trials were analyzed to determinewhether there was evidence for short-term memory during bursting trials, noevidence for short-term memory wasobserved. Figure 13 shows the burstingtrial responses from the same neuronalnetworks as Figure 12. These results in-dicate that, on trials where a networkburst occurs, the memory of the cuestimulus is lost and the response to theprobe stimulus does not reflect cue-related information. For both dishes,network bursts disrupted hidden mem-ory for the cue.

DiscussionWe have employed a combination of op-togenetics, MEA recordings, and machineclassifiers to investigate how stimulus in-

formation is encoded and stored in networks formed by dissoci-ated cortical neurons in culture. The results from ourexperiments indicate that stimulus-specific information is main-tained for durations longer than 1 s and that both active andhidden memory processes are present.

Figure 11. Encoding of short-term memory and test for two-back memory. A, Time series charting cue-related informationduring an experiment with a DIV11 culture. Cues are presented for 150 ms, the delay is 350 ms, and the probe is presented for 150ms. All stimuli are random dot patterns. Hyphenated black line indicates accuracy of classification of testing data by LSM. Thin solidblue line indicates classification accuracy when temporal information is filtered out. Solid green line indicates classification accu-racy when spatial information is eliminated by pooling spikes. Dotted red line indicates classification rate based only on overallspike count. Data indicate that most stimulus information is encoded in the pattern of active channels, not spike counts or timing.Chance rate of classification is 0.25 (solid gray line); other conventions as in Figure 7. B, Response of DIV7 dish to cue-probe-probetrials. A “one-back” memory was observed, where the probe presentation elicited an increase in the amount of cue-relatedinformation. However, presentation of the second probe did not lead to a similar recovery of cue-related information. “A” and “B”are the two cue stimuli to be classified. Either “C” or “D” was randomly assigned to be the first probe stimulus. The graphs abovepresent the average across these two cases, indicating that in these two cases a short-term memory could be detected. No“two-back” memory for cue identity was detected. Top, Time series showing available cue-related information. Bottom, Spikerates. Solid line indicates the chance rate of classification for the cue stimulus (50%); and the dotted lines indicate the 95%confidence intervals for a random label reassignment. Other conventions as in Figure 6.

Dranias et al. • Short-Term Memory J. Neurosci., January 30, 2013 • 33(5):1940 –1953 • 1949

These experiments used a new optoge-netic approach. Electrical stimulationelicits an artifact at the time of stimulationfollowed by a highly stereotyped, fixed-order response lasting 10 –30 ms. The first5–15 ms of this initial response is artifac-tual in that it appears to be mediated elec-trically, not synaptically (Kermany et al.,2010). A synchronized network burst of-ten follows this initial response with apeak occurring between 20 and 200 ms,before a return to the prestimulus state(Jimbo et al., 1999; Eytan and Marom,2006; Shahaf et al., 2008). With opticalstimulation, there is no stimulus artifact,and the latency to the first recorded spikeis much longer, between 50 and 200 ms.The initial response to optical stimulationtypically involves fewer neurons and alsois fairly stereotyped (Fig. 5). As with elec-trical stimulation, a network burst is usu-ally induced by optical stimulation, but ata much later latency (typically in the rangeof 100 –300 ms) and for a longer duration,typically lasting 300 –500 ms before activ-ity dissipates to prestimulation levels.

On both bursting and nonbursting tri-als, most actively encoded stimulus infor-mation is found in the first 200 ms. This isalso a period where the network responseis less variable. During this initial 200 ms,stimulus information is primarily en-coded by the spatial configuration of ac-tive units. However, this is not a simple,static spatial pattern of units that resem-bles the image of the optical stimulus; only50% of units respond to direct illumina-tion and the units that encode the stimu-lus change over time, with the relationshipbetween network activity and stimulusidentity gradually deteriorating into ran-domness. On bursting trials, actively en-coded stimulus information can persist aslong as 1 s, but this information appears tobe mainly encoded by differences in over-all spike counts or temporal information,particularly differences in the onset andtermination of network bursts.

When the network was challengedwith a second “probe” stimulus, whichcontains no information regarding thefirst “cue” stimulus, the response to theprobe was typical and primarily encodedin the spatial pattern of active units. How-ever, there was evidence for a residualmemory of the previously presented cuestimulus. The mechanisms that encodethis short-term memory were not singu-lar. In two dishes, the ability of the probeto recall cue information relied on the ab-sence or presence of a network burst, andthe short-term memory was primarily en-coded by a change in overall spike count.

Figure 12. Hidden memory encodes stimulus identity in overall spike counts. A, Raster plot of sorted nonbursting cue-probetrials from same dish as Figure 6; 50 trials of data are shown. Conventions are as in Figure 7. B, Encoding of stimulus information.Both cue and hidden memory are primarily encoded in the overall spike rate (dotted red line). Filtering out of temporal informationhas no effect (solid blue line), and filtering out of single unit activity by pooling spikes has little effect (solid green line). Chanceclassification performance is 50% (gray line); classification of unfiltered nonbursting trials is shown as hyphenated black line. C,Raster plot of sorted nonbursting data from cue-probe trials recorded from a different DIV7 MEA. D, As with the other dish, themajority of stimulus information recorded in hidden memory is encoded in overall spike counts (dotted red line). Disruptingtemporal (blue) or spatial (green) information fails to reduce classifier accuracy from control levels (hyphenated black line). Otherconventions are as in Figure 10.

1950 • J. Neurosci., January 30, 2013 • 33(5):1940 –1953 Dranias et al. • Short-Term Memory

In another dish, short-term memory wasmainly encoded by changes in the spatialpattern of active units.

Important questions concerning theinformation content of synchronous net-work bursts were also answered. Networkbursts are capable of representing and ex-tending stimulus information in activememory, but simultaneously it was shownthat network bursts erased hidden mem-ories of the cue. If hidden memory playsan important role in short-term memory,then these observations may have someimplications for cognitive impairmentsthat are associated with pathological dis-charges that can accompany seizure disor-ders. This is an area that requires moreinvestigation.

The observation that a probe stimulus canelicit information about previously presentedstimuli compares favorably with some the-oretical models of short-term memory.Mongillo et al. (2008) proposed a modelof working memory wherein a randomlyorganized neural network with adaptivesynapses is presented with a stimulus andthen periodically maintains the stimulusinformation by reactivating the networkwith a uniform or nonspecific probe in-put. Figure 9A shows one such case wherewe used a uniform white stimulus as theprobe and a observed a probe-dependentextension of cue-related informationHowever, whenever we administered asecond probe stimulus, we were unableobserve a “two-back” memory effect (Fig.11B). The theory may, however, antici-pate this result as one of its primary prem-ises is that excitatory synapses inprefrontal cortex are facilitating. The ex-citatory neurons in our living neuronalnetworks are likely to mainly have de-

4

Figure 13. No short-term memory is observed on burstingtrials. A, Raster plot of sorted bursting cue-probe trials fromsame dish as Figure 6; 50 trials of data are shown. Conventionsare as in Figure 7. B, Encoding of stimulus information. No“short-term” memory is seen after probe presentation. Cueinformation is primarily encoded in the overall spike rate onbursting trials (dotted red line). Filtering out of temporal infor-mation has no effect (solid blue line), and filtering out of singleunit activity by pooling spikes has little effect (solid green line).Chance classification performance is 50% (gray line); classifi-cation of unfiltered nonbursting trials is shown as hyphenatedblack line. C, Raster plot of sorted bursting data from cue-probe trials recorded from a different DIV9 MEA. D, As with theother dish, no short-term memory is seen on bursting trials.For this DIV9 MEA, cue information is mostly encoded tempo-rally (blue and red). The loss of information is less when onlyspatial information is eliminated before classification (green).Chance classification rate is 50%. Other conventions are as inFigure 10.

Dranias et al. • Short-Term Memory J. Neurosci., January 30, 2013 • 33(5):1940 –1953 • 1951

pressing synapses, such as rat sensory cortex (Markram et al.,1998).

The results from our cue-probe experiment more closelymatch the results reported by Nikolic et al. (2009), who recordedfrom the visual cortex of anesthetized cats. Despite the large dif-ferences between our in vitro system and the in vivo system usedby Nikolic et al. (2009) there are many similarities. The durationof memory for single stimuli is comparable with fading stimulusmemories reported as lasting �700 ms in vivo while in vitro themean was 675 ms. The correlation between mean firing rate andcorrect classification rates is about the same in vivo (r � 0.65) andin vitro (r � 0.64), indicating that mean firing rate carries a sim-ilar amount of information in each system. The presence of a“one-back” memory but absence of a “two-back” memory wasalso reported in vivo. Nikolic et al. (2009) hypothesized that theabsence of this “two-back” memory might be a mechanism asso-ciated with visual masking, but the fact that this trend exists invitro suggests that it is a generic property of neural networksrather than an unusual property of visual cortex. The networksalso have their differences. The in vitro network is prone to net-work bursts after �300 ms. The ratio of inhibitory to excitatoryneurons is similar for the two systems (Marom and Shahaf,2002), so the added stability of the in vivo system likely is a func-tion of its architecture. Another cause for this difference may bedevelopmental rather than architectural, reflecting the absence ofexternal inputs to the in vitro network during development.When networks growing on an MEA were continuously stimu-lated with random inputs, the development of synchronized net-work bursting was suppressed (Wagenaar et al., 2006b). Hence,external signals may be needed during development to ensurenetwork stability, by helping to tune and balance the outputs ofexcitatory and inhibitory neurons.

The memory processes we demonstrate endow the networkwith the ability to perform state-dependent computations, suchthat the response to a stimulus varies with the identity of thepreceding stimulus. This constitutes a form of “reservoir” com-puting (Buonomano and Maass, 2009): nonlinear interactions intime and space. The usual application for reservoir computing ispattern separation, and these nonlinear interactions allow similarsignals to evolve into very different network states. The short-term memory observed during the cue-probe presentation ex-periments demonstrates that networks of disassociated corticalneurons can have a nonlinear sensitivity to temporal order. At thesame time, our time-series analysis of the network response tooptical stimulation demonstrates that these networks have a fad-ing memory that preserves stimulus information long enough toallow recurrent interactions. The only complication is that net-work bursts may limit some computations; although they canpropagate stimulus information they appear to erase hiddenmemory.

In conclusion, optogenetic methods can allow researchers toexplore neuronal mechanisms of memory in vitro using stimuliand experimental designs similar to those used in vivo. An in vitroplatform provides the additional benefit of allowing researchersto rapidly screen for drugs that might have therapeutic effects,manipulate genes and gene expression, provide a test bed for stemcell research, or examine how computers might be used in con-junction with brain cells to enhance memory and informationprocessing (Chiappalone et al., 2003; Bakkum et al., 2008a; Jainand Muthuswamy, 2008; Berger et al., 2011; Stephens et al.,2012). All these possibilities can have far reaching consequences

for the treatment of diseases associated with short-term memoryimpairments.

ReferencesBaddeley A (1996) The fractionation of working memory. Proc Natl Acad

Sci U S A 93:13468 –13472. CrossRef MedlineBakkum DJ, Chao ZC, Potter SM (2008a) Spatio-temporal electrical stimuli

shape behavior of an embodied cortical network in a goal-directed learn-ing task. J Neural Eng 5:310 –323. CrossRef Medline

Bakkum DJ, Chao ZC, Potter SM (2008b) Long-term activity-dependentplasticity of action potential propagation delay and amplitude in corticalnetworks. PloS One 3:e2088. CrossRef Medline

Beggs JM, Plenz D (2003) Neuronal avalanches in neocortical circuits.J Neurosci 23:11167–11177. CrossRef Medline

Berger TW, Hampson RE, Song D, Goonawardena A, Marmarelis VZ, Dead-wyler SA (2011) A cortical neural prosthesis for restoring and enhancingmemory. J Neural Eng 8:046017. CrossRef Medline

Berndt A, Schoenenberger P, Mattis J, Tye KM, Deisseroth K, Hegemann P,Oertner TG (2011) High-efficiency channelrhodopsins for fast neuronalstimulation at low light levels. Proc Natl Acad Sci U S A 108:7595–7600.CrossRef Medline

Buonomano DV, Maass W (2009) State-dependent computations: spa-tiotemporal processing in cortical networks. Nat Rev Neurosci 10:113–125. CrossRef Medline

Chang C-C, Lin C-J(2011) LIBSVM: a library for support vector machines.ACM Trans Intell Syst Technol 2:1–27.

Chiappalone M, Vato A, Tedesco MB, Marcoli M, Davide F, Martinoia S(2003) Networks of neurons coupled to microelectrode arrays: a neuro-nal sensory system for pharmacological applications. Biosens Bioelectron18:627– 634. CrossRef Medline

Dockendorf KP, Park I, He P, Príncipe JC, DeMarse TB (2009) Liquid statemachines and cultured cortical networks: the separation property. Bio-systems 95:90 –97. CrossRef Medline

Dranias MR, Ju H, Rajaram E, VanDongen AM (2011) Spatiotemporal andfading memory properties in an in vitro model of information processing.Soci Neurosci Abstr 37:239.206.

Eytan D, Marom S (2006) Dynamics and effective topology underlying syn-chronization in networks of cortical neurons. J Neurosci 26:8465– 8476.CrossRef Medline

Eytan D, Brenner N, Marom S (2003) Selective adaptation in networks ofcortical neurons. J Neurosci 23:9349 –9356. Medline

Gal A, Eytan D, Wallach A, Sandler M, Schiller J, Marom S (2010) Dynamicsof excitability over extended timescales in cultured cortical neurons.J Neurosci 30:16332–16342. CrossRef Medline

Haykin SS (1999) Neural networks: a comprehensive foundation, Ed 2. Up-per Saddle River, NJ: Prentice Hall.

Jain T, Muthuswamy J (2008) Microelectrode array (MEA) platform fortargeted neuronal transfection and recording. IEEE Trans Biomed Eng55:827– 832. CrossRef Medline

Jimbo Y, Tateno T, Robinson HP (1999) Simultaneous induction ofpathway-specific potentiation and depression in networks of cortical neu-rons. Biophys J 76:670 – 678. CrossRef Medline

Kermany E, Gal A, Lyakhov V, Meir R, Marom S, Eytan D (2010) Tradeoffsand constraints on neural representation in networks of cortical neurons.J Neurosci 30:9588 –9596. CrossRef Medline

Kobayashi R, Tsubo Y, Shinomoto S (2009) Made-to-order spiking neuronmodel equipped with a multi-timescale adaptive threshold. Front Com-put Neurosci 3:9. CrossRef Medline

Maass W, Natschlager T, Markram H (2002) Real-time computing withoutstable states: a new framework for neural computation based on pertur-bations. Neural Comput 14:2531–2560. CrossRef Medline

Maheswaranathan N, Ferrari S, Vandongen AM, Henriquez CS (2012)Emergent bursting and synchrony in computer simulations of neuronalcultures. Front Comput Neurosci 6:15. CrossRef Medline

Markram H, Wang Y, Tsodyks M (1998) Differential signaling via the sameaxon of neocortical pyramidal neurons. Proc Natl Acad Sci U S A 95:5323–5328. CrossRef Medline

Marom S, Shahaf G (2002) Development, learning and memory in largerandom networks of cortical neurons: lessons beyond anatomy. Q RevBiophys 35:63– 87. CrossRef Medline

Mongillo G, Barak O, Tsodyks M (2008) Synaptic theory of working mem-ory. Science 319:1543–1546. CrossRef Medline

1952 • J. Neurosci., January 30, 2013 • 33(5):1940 –1953 Dranias et al. • Short-Term Memory

Nagel G, Brauner M, Liewald JF, Adeishvili N, Bamberg E, Gottschalk A(2005) Light activation of channelrhodopsin-2 in excitable cells of Cae-norhabditis elegans triggers rapid behavioral responses. Curr Biol 15:2279 –2284. CrossRef Medline

Natschlager T, Markram H (2003) Computer models and analysis tools forneural microcicuits. In: Neuoscience databases: a practical guide, 2003edition (Kotter R, ed), pp 123–138. Norwell, MA: Kluwer Academic.

Nikolic D, Hausler S, Singer W, Maass W (2009) Distributed fading mem-ory for stimulus properties in the primary visual cortex. PLoS Biol7:e1000260. CrossRef Medline

Schrader S, Grun S, Diesmann M, Gerstein GL (2008) Detecting synfirechain activity using massively parallel spike train recording. J Neuro-physiol 100:2165–2176. CrossRef Medline

Shahaf G, Eytan D, Gal A, Kermany E, Lyakhov V, Zrenner C, Marom S(2008) Order-based representation in random networks of cortical neu-rons. PLoS Comput Biol 4:e1000228. CrossRef Medline

Stephens CL, Toda H, Palmer TD, DeMarse TB, Ormerod BK (2012) Adultneural progenitor cells reactivate superbursting in mature neural net-works. Exp Neurol 234:20 –30. CrossRef Medline

Van de Ven TJ, VanDongen HM, VanDongen AM (2005) The nonkinasephorbol ester receptor � 1-chimerin binds the NMDA receptor NR2Asubunit and regulates dendritic spine density. J Neurosci 25:9488 –9496.CrossRef Medline

Wagenaar DA, Nadasdy Z, Potter SM (2006a) Persistent dynamic attractorsin activity patterns of cultured neuronal networks. Phys Rev E Stat NonlinSoft Matter Phys 73:051907. CrossRef Medline

Wagenaar DA, Pine J, Potter SM (2006b) Searching for plasticity in dissoci-ated cortical cultures on multi-electrode arrays. J Negat Results Biomed5:16. CrossRef Medline

Wang XJ (2001) Synaptic reverberation underlying mnemonic persistentactivity. Trends Neurosci 24:455– 463. CrossRef Medline

Dranias et al. • Short-Term Memory J. Neurosci., January 30, 2013 • 33(5):1940 –1953 • 1953